bhack
bhack
Any update?
@agunapal I am adding `g++` in runtime images with https://github.com/pytorch/pytorch/pull/138612 to solve https://github.com/pytorch/pytorch/issues/116696 (`torch.compile`) but I suppose that it is not enough for you export. What do you think?
On pytorch ecosystem Kornia has this package organizzation: https://kornia.readthedocs.io/en/latest/
It Is under Alphabet so I think that you can make an internal pass before investing time and create a duplicate.
After solving this another one appeared around the corner https://github.com/pytorch/pytorch/issues/139440
I suppose that we need to achieve this with retraining/fine-tuning. What is the plan?
It could be interesting to take a look at the approach in: https://github.com/openxla/openxla-pjrt-plugin
> @LukeWood > >so we'll have to roll a pure TF version > > Could you clarify? Do you mean this should be a custom C++ op, similar to `PairwiseIoUOp`?...
Thanks @cantonios. This is also true for other ops in `tf.image` namespace. I suppose the problem is more related to the "barriers" between projects. When an issue emerge downstream but...
P.s. @cantonios Just in case you are interested we have a similar issue with the MLIR bridge path on `tf.ImageProjectiveTransformV3` (https://github.com/tensorflow/tensorflow/pull/55335/files#r832749400). What is not clear is if instead of rewriting...